>>7
Shaders are slower than software rendering? Are you brain dead?
Shaders are software. They allow you to program a part of the rendering pipeline. They usually use a restricted subset of C with some additional extensions.
Compute kernels have evolved from shaders, and provide a much richer subset of general purpose programming languages. Compute kernels are usually C or C++ based depending on what technology you're using.
Shaders and compute kernels are software designed to run in parallel on heterogenous computing devices like GPUs.
The only reason software rendering is making a return is because heterogenous computing languages like OpenCL/DirectCompute/C++ AMP are making GPGPU programming much more accessible and familiar feeling. That's right. The future of software rendering still won't be done on traditional, out-of-order, cache-coherent general purpose CPUs. It's being done on heterogeneous devices--GPUS--and with new languages designed for extreme levels of parallelization. And no, things like Go, Scala, Haskell or even Clojure are not designed for extreme levels of parallelization, they're just some experiment that ventured part way into parallel computing and ended up with a half-baked result.
I won't classify CUDA in the same league as the aforementioned languages, mainly because it's shit and still doesn't implement half the features in DirectCompute or OpenCL 1.1. nVidia still hasn't released full OpenCL 1.1 drivers yet either. Fuck nVidia and their proprietary bullshit.